GPT-5.6 Sol, Terra & Luna: Pricing, Benchmarks, and How It Compares to Fable 5 (2026)

What is GPT-5.6?

GPT-5.6 is OpenAI's model family released on July 9, 2026, in three named tiers: Sol (the flagship), Terra (balanced, for everyday work), and Luna (fast and cheap). In the new naming system the number is the generation and the names are durable capability tiers that can advance on their own cadence. All three carry a roughly 1.05M-token context window and up to 128K output tokens, and it's generally available across ChatGPT, Codex, and the OpenAI API.

Two things to know up front. The launch was a staged, regulated rollout (early access was gated to limited preview partners before the wider release), and inside OpenAI's tools the Codex app is now folded into ChatGPT — "ChatGPT Codex" — so the branding shifted even though the coding agent is the same. For the APIs behind this and rival models, see our best AI APIs for developers guide.

Sol vs Terra vs Luna: which tier should you use?

Pick the tier by the job, not the label. Sol is for the hardest reasoning and agentic coding; Terra is the everyday workhorse at a lower price; Luna is for high-volume, latency-sensitive work where cost matters most.

How much does GPT-5.6 cost?

Sol runs $5 per million input tokens and $30 per million output; Terra is $2.50 / $15; Luna is $1 / $6. That makes Sol roughly half the price of Claude Fable 5's single $10 / $50 tier on output — a real gap if you're generating a lot. GPT-5.6 also adds more predictable prompt caching (explicit cache breakpoints, a 30-minute minimum cache life; cache writes bill at 1.25× the uncached input rate, reads keep the ~90% discount), which cuts cost further on repeated context.

GPT-5.6 vs Claude Fable 5: what the benchmarks show

This is where the launch-day "it's better than everything" claims meet the data — and the data splits. On the independent Artificial Analysis Coding Agent Index, Sol leads at 80, with Terra at 77.4 and Claude Fable 5 at 77.2 (between Terra and Luna). Sol also tops Terminal-Bench 2.1 (agentic, terminal-driven coding) at 88.8%, ahead of Fable 5 around 84% and Claude Opus 4.8 at 78.9%.

But Fable 5 still wins the test closest to real repos

Flip to SWE-bench Pro — end-to-end resolution of real GitHub issues — and Claude takes it clearly. Fable 5 posts 80.3%, Opus 4.8 69.2%, and GPT-5.5 58.6%; OpenAI has not published an official SWE-bench Pro score for any GPT-5.6 tier, and third-party leaderboard tracking places Sol around 64.6% — well behind Fable 5. On the composite Artificial Analysis Intelligence Index the two are a statistical tie (Fable 5 at 60, Sol at 59), though Artificial Analysis runs Fable 5 with an Opus 4.8 fallback.

So which is better — GPT-5.6 Sol or Fable 5?

It depends on the work, and the split is consistent. Reach for Sol when you're driving an agent harness or terminal-based coding, want fast, token-efficient output, and care about cost — it leads the agentic benchmarks and is about half Fable 5's price. Reach for Claude (Fable 5, or Opus 4.8) for hard, real-world repository work, where SWE-bench Pro says it still resolves more issues correctly.

Put simply: agent-harness and terminal coding lean Sol; real repository work still leans Claude. Many teams will run both — Sol as the fast, cheap daily driver and orchestrator, Claude for the gnarly changes — which is exactly what an aggregator or a multi-model setup makes easy. We compare the two leading coding agents in Codex vs Claude Code, and the whole field in the best AI coding tools.

What's new versus GPT-5.5?

The headline change is the tiered family itself — one generation, three durable capability levels — which replaces the older single-model-plus-mini split and gives a cleaner choice across intelligence, speed, and cost. On top of that: a materially stronger agentic-coding result (Sol tops the Coding Agent Index and Terminal-Bench, where GPT-5.5 trailed), the more predictable prompt caching above, and a robust safety stack that OpenAI stress-tested for weeks before the regulated rollout. GPT-5.5 remains available for anyone who doesn't want to migrate yet.

A vibe-coder's first look (early-user notes)

Beyond the benchmarks, here's what early users report after driving Sol in real agent setups — treat it as hands-on impression, not measurement. The recurring praise: Sol is a notably strong orchestrator. In multi-agent harnesses it will spin up and prompt subagents on its own — the first model several practitioners say does that without being forced — and the sub-prompts it writes are markedly better than GPT-5.5's, closer to a planning model. Some found it a better orchestrator than Opus 4.8 in their own workflow.

The caveats are just as consistent: its front-end design output is improved but still reads as "AI slop" (a newer style, but slop), and it's weak at complex generative visuals — one hands-on test of a 3D solar-system visualization went poorly, where a Claude model did better. Net of the hands-on reports and the benchmarks: Sol is a fast, cheap, self-orchestrating daily driver that's genuinely strong at agentic coding, with real limits on design and hardest-case repo work.

Where can you use GPT-5.6?

All three tiers are available through the OpenAI API. Inside OpenAI's own tools, Sol is in ChatGPT and in Codex (now "ChatGPT Codex"); Terra and Luna are reached mainly via the API. Access to Sol has been gated to higher paid plans rather than the entry tier, and it's already wired into third-party tools — Cursor and others added it on launch day. For building on it directly, the API pricing and tiers above are what matter; for a ready-made coding agent, see the best AI coding tools and Claude Code or ChatGPT in our directory.